FutureX · Physical AI Daily — Issue 57 (07/14)

The US NHTSA has issued an ultimatum to robotaxi operators to submit remediation plans for emergency-response obstruction by month's end. China Mobile procured 400 humanoid robots for 124 million yuan, with Zhiyuan and Unitree receiving major shares. Galbot won a 236 million yuan bid for 500 robots at Yibin high-speed rail station. ByteDance is exploring autonomous driving led by its Seed world model team, though the company denies business plans. South Korea's Holiday Robotics raised 155 billion won in a record Series A. MIT researchers proposed B-spline Policy to speed up manipulation policy inference by parameterizing actions as continuous curves. A new method achieves 5x sample efficiency for VLA post-training via hindsight relabeling. BeyondSight restores object permanence in end-to-end autonomous driving, and PanoWorld uses panoramic rotational equivariance to solve world models' long-horizon memory problem.

Today's Highlights · US NHTSA issues ultimatum to all robotaxi operators: submit remediation plans for emergency-response obstruction by month's end · China Mobile procures 400 humanoid robots for 124 million yuan , with Zhiyuan Chinese humanoid robotics startup getting 78 million and Unitree getting 46.05 million · Galbot Chinese embodied AI startup wins 236 million yuan bid for Yibin high-speed rail station, 500 robots setting a single-procurement record · ByteDance explores autonomous driving, led by its Seed world model team ; company officially denies any business plans · South Korea's Holiday Robotics raises 155 billion won , setting the country's Series A funding record B-spline Policy: Parameterizing actions as continuous curves speeds up manipulation policy inference · manipulation One real-time bottleneck for manipulation policies is predicting discrete-time action chunks step by step, with compute stacking linearly with the number of steps. MIT's Antonio Torralba and colleagues propose B-spline Policy BSP : instead of predicting a discrete action sequence, it directly predicts the parameters of a continuous B-spline curve defined by knots and control points, with actions sampled from the curve. This representation can be seamlessly integrated into existing policy learning pipelines diffusion policy, ACT, etc. can all be adapted , significantly cutting inference overhead while maintaining success rates — moving the "speedup" problem from model compression to the action representation layer. Xiaoshen Han et al. MIT et al. · arXiv 2607.09648 source https://arxiv.org/abs/2607.09648 Learning from Hindsight: 5x sample efficiency gain for VLA post-training · vla The biggest cost of RL post-training for VLAs is real-robot rollouts — every policy update burns expensive, slow real-world interaction data, and failed rollouts are typically discarded outright. This paper introduces hindsight relabeling into VLA RL post-training: failed rollouts aren't wasted — they're re-scored and reused based on what task they actually accomplished . The result is a 5x improvement in sample efficiency, outperforming dense progress-reward baselines on LIBERO's out-of-distribution tasks. Iris Xu et al. · arXiv 2607.09042 source https://arxiv.org/abs/2607.09042 BeyondSight: Restoring "object permanence" to end-to-end autonomous driving · autonomy Most end-to-end driving systems implicitly bind "whether a target exists" to "whether it is currently observed" — a vehicle occluded by another car for an extended period gradually fades and even disappears from the system's view, causing downstream prediction and planning to drop critical traffic participants. BeyondSight, from University of Toronto's Waslander group, decouples existence from observability, maintaining persistent object hypotheses across time; it also releases nuScenes-Permanence , providing observability-conditioned supervision and benchmarks for driving evaluation under long-duration occlusion. Sandro Papais et al. University of Toronto · arXiv 2607.09138 source https://arxiv.org/abs/2607.09138 PanoWorld: Using panoramic rotational equivariance to solve world models' long-horizon memory problem · world-model · HF↑7 Video world models tend to "forget" places they've already traversed after running for a while — when looking back, the scene has already drifted. PanoWorld exploits the rotational equivariance property of omnidirectional representations: treating rotation as an implicit geometric transform, thereby reducing camera trajectories to translation under a fixed orientation, then using Dense Panoramic Ray-Conditioning with geometry-aware memory augmentation to maintain long-horizon consistency. HF community score ↑7 , the most-discussed world model paper today. Haoyuan Li et al. · arXiv 2607.09661 source https://arxiv.org/abs/2607.09661 CD-LAM: World models' "action controllability" turns out to be contaminated by data bias · world-model Action-conditioned world models ACWMs need to learn to "predict future frames given an action," but action-labeled data is expensive, so the field commonly uses latent action models LAMs to infer actions back from unlabeled video — the problem is that latent actions learned this way absorb large amounts of spurious correlation unrelated to the actual action. CD-LAM restructures this pipeline from a causal debiasing perspective, comprehensively improving latent action controllability, downstream action following, and visual fidelity, while requiring only 6k fine-tuning steps and reducing real-robot action adaptation updates by more than 12x . Yufan Wei et al. · arXiv 2607.09185 source https://arxiv.org/abs/2607.09185 TactiDex: Dexterous manipulation should be evaluated on contact, not on how human-like the motion looks · benchmark Current human-to-robot hand transfer pipelines mainly replicate kinematic trajectories, resulting in "motion imitation" rather than "physical contact" — the gesture looks human, but contact force and grasp stability don't match up. TactiDex is a real-robot, tactile-guided evaluation benchmark that shifts the evaluation criterion from kinematic similarity to contact-level human-likeness ; its companion TactiSkill framework uses a three-component tactile reward, treating tactile signals as structured learning signals rather than post-hoc verification. As dexterous hands move into mass production, tightening evaluation standards matters more than adding a few more degrees of freedom. Suting Ni et al. · arXiv 2607.09190 source https://arxiv.org/abs/2607.09190 The energy bottleneck for edge VLMs isn't "seeing" — it's "talking" · benchmark VLMs are the perception backbone of embodied AI, and industry efficiency optimization has almost entirely bet on cutting visual tokens, assuming "looking at images" is the default power drain. This paper conducts the first systematic energy profiling of on-device VLM inference, with a counterintuitive conclusion: the real energy bottleneck is how many words the model outputs , not how many pixels it looks at. For teams trying to fit VLAs onto robot bodies with tight compute and battery budgets, this directly reshapes optimization priorities. Junfei Zhan et al. · arXiv 2607.09520 source https://arxiv.org/abs/2607.09520 VLANeXt: A recipe study for building strong VLA models · vla Accepted to ICML 2026. Shanghai Jiao Tong University and Nanyang Technological University systematically ablate VLA training recipes, yielding several directly actionable engineering findings — the most counterintuitive being that connecting the VLM and policy module via soft connection slightly outperforms the two common extremes of loose coupling and tight coupling. For teams building their own VLA architectures, this kind of recipe-level finding is more useful than yet another SOTA number. SJTU × NTU · ICML 2026 · Analysis: Ru-hulu's Embodied AI Sharing Session source WeChat, CN https://mp.weixin.qq.com/s? biz=MzY4MzM0MjYzOQ==&mid=2247483827&idx=1&sn=e94696785030458a8720dde26f33a5d3&chksm=f2ab9d22ffdf10defe5f7b9b21de70fddd08f932234fd6df593696fa0baca96f3f8a5862d58c rd Other papers today: FlowDAgger uses action inversion to map human expert corrections back into the noise space of the base policy, enabling online human-robot collaborative adaptation for generative policies arXiv 2607.08877 source https://arxiv.org/abs/2607.08877 ; GenVid2Robot uses rigid-body geometric consistency to convert generated video into executable real-robot trajectories arXiv 2607.09191 source https://arxiv.org/abs/2607.09191 ; PhysV2A uses reachability gating plus semantic mask constraints to complete video-to-action feasibility arXiv 2607.09365 source https://arxiv.org/abs/2607.09365 ; AgenticFocus synthesizes ordinary first-person human video via mixed reality into humanoid-trainable demonstration data arXiv 2607.08857 source https://arxiv.org/abs/2607.08857 ; DemoBridge converts monocular RGB stereo recordings into physically-verified robotic arm trajectories arXiv 2607.09519 source https://arxiv.org/abs/2607.09519 ; CLAP prepends a natural-language action description before action tokens, enabling lossless VLM-to-VLA adaptation arXiv 2607.08974 source https://arxiv.org/abs/2607.08974 ; OpenLongTail open-sources a generative data engine for long-tail driving scenarios arXiv 2607.09655 source https://arxiv.org/abs/2607.09655 ; 4DR360 builds semantic occupancy into persistent scene state for full-scene fusion of 4D mmWave radar and cameras arXiv 2607.09629 source https://arxiv.org/abs/2607.09629 ; CORAL-AUV deploys, for the first time, a zero-shot RL policy trained on a CFD-surrogate drag model onto a 6-DoF underwater robot arXiv 2607.09557 source https://arxiv.org/abs/2607.09557 ; Hydra++ builds hierarchical 3D scene graphs in real time with object-level shape estimation, supporting hybrid lidar-camera configurations arXiv 2607.09455 source https://arxiv.org/abs/2607.09455 . · Amap Alibaba mapping unit ABot-World Studio : given an image, a sentence, or a video clip, it generates a freely walkable 3D world locally, which can be solidified into a shareable 3D scene. The key breakthrough is two removed barriers — inference runs on a single consumer-grade RTX 5090-class GPU rather than a cloud cluster, and based on the LongForcing training paradigm, it has been verified to generate for 24 continuous hours without noticeable drift previous interactive world models typically stayed stable for only a few minutes, with the world vanishing once interaction ended . Code and model weights have been simultaneously open-sourced to GitHub, ModelScope, and Hugging Face source WeChat, CN https://mp.weixin.qq.com/s? biz=MzE5ODQ3MDM4Mw==&mid=2247491473&idx=1&sn=5b5bd13dfb21a9f41e3333d7fb8e0e54&chksm=97dd585d771ea4375f6e54bddad150739e6da09ddb4462adda0903432175b043a74e48e74e69 rd · DexJoco : a MuJoCo-based dexterous hand manipulation evaluation benchmark, paired with a low-cost data collection scheme — using Rokoko motion-capture gloves to capture finger poses, retargeted to Allegro hand joints via GeoRT source WeChat, CN https://mp.weixin.qq.com/s? biz=MzAwMTY3NjA1OA==&mid=2247489517&idx=2&sn=e4abfe34c9e0ac2b6e91f094ce41ab5f&chksm=9b834e01c0a166d4a0e7b29a1f5bfe38534f7737581e1d42703a46145a1b19b550c50b4e599b rd · Zhiyuan LinkSoul Community : a robot interaction agent-building platform launches, with wake words, persona, voice tone, knowledge base, skills, user recognition, and other configuration items all made into a visual configuration interface source WeChat, CN https://mp.weixin.qq.com/s? biz=MzE5ODg0NTY2Mw==&mid=2247485303&idx=1&sn=09939b28a19b3afaa0ddfefd0d5e193e&chksm=97b37fce10de78b10f001ab309557eff2df2c88440b8f46a2c81963e683c49e99479cdcd77a7 rd Holiday Robotics South Korea | Series A | 155 billion won ~$110M | Valuation 700 billion won · humanoid IMM Investment led, with existing shareholders Stonebridge and Atinum adding on, and SL Investment and KB Investment joining as new investors — most notably, the Korea Development Bank and the Industrial Bank of Korea, two policy-oriented financial institutions, both participated directly in this Series A , marking direct state industrial-capital backing for a startup. This sets South Korea's Series A funding record. The company follows an "industrial-first, wheeled-deployment" path: its core product FRIDAY stands 176cm tall, weighs 115kg, and has 64 degrees of freedom across the body, using a 66kg heavy-duty chassis to lower its center of gravity and supporting hot-swappable batteries for continuous operation. Its production timeline is aggressive — commercial delivery in H2 2026 with 100 units for the full year, scaling to 1,000 units/year in 2027, with a long-term goal of pushing the per-unit price down to 100 million won ~$70,000 . Founder Song Ki-young is a serial entrepreneur in industrial AI vision formerly of Sualab , not from an academic lab background. ⚠️ Production timeline per company planSource: Embodied Emergence source WeChat, CN https://mp.weixin.qq.com/s? biz=MzYzMzYwNTEwMg==&mid=2247493131&idx=2&sn=e9f4537ba076a06671766b9e42d91f1a&chksm=f1d67e7e23f035341f2aa98692cca0ef5611ab541fea25b28684920616e6efd6e29628f271a1 rd Ling Cha Yun Kong Shenzhen | Series C++ | Several hundred million yuan · hardware Led by CDF Capital, with Guotai Junan Innovation Investment participating, and existing shareholder Huakong Fund adding on. This company makes integrated joint modules and magnetic encoders; as of June 2026, it has shipped over 50,000 joint module units cumulatively to more than 2,000 customers, making it China's largest shipper in this niche. Technically, it uses a dual-encoder, fully closed-loop approach one absolute-value encoder at the motor end and one at the reducer output end , achieving repeat positioning accuracy of ±7 arcseconds. Humanoid robots already contribute over 30% of revenue , and the company's 2025 revenue doubled year-over-year. This round's funding will mainly go toward capacity expansion — targeting an increase from tens of thousands of units to over 100,000 units in annual capacity. Source: Embodied Yuan source WeChat, CN https://mp.weixin.qq.com/s? biz=MzYzOTY0NDk5MA==&mid=2247487679&idx=2&sn=255198dd3569f117633dc981d34fd079&chksm=f10363f53b24359ebcf0d4a5e4f9b16d1fc7be8b53ccc5fd44cfa260d4c44d4d56647c3e1b7e rd Qingyan Precision | B2, B3 series | Several hundred million yuan · hardware Led by the SASAC Equipment Manufacturing Industry Fund , with six automakers jointly adding on — automakers collectively betting on a components supplier is a frequently recurring signal in the embodied AI upstream supply chain this year. Source: 36Kr source https://m.36kr.com/p/3893535788890760 Luobo Intelligent Robopoet | Pre-A round | Tens of millions of yuan class · adjacent Co-led by CMC Capital and Fibocom, with Tuya Smart participating, and existing shareholders Sequoia China and GSR Ventures adding on. Makes companion robots, whose products have previously sold through Huawei's retail channels. There's another reference point in the companion-robot space today — market prices for bionic companion robots range from 15,800 to 990,000 yuan , a 60x spread, showing the product category is far from converging on a definition. Source: Robot Outlook source WeChat, CN https://mp.weixin.qq.com/s? biz=Mzk0NDU5NjMwMA==&mid=2247492833&idx=1&sn=42a212484408978169e5ebaa6f02f680&chksm=c24f800e044c3a5c1dd379c902ee7c52d317d840a5343a52e954496a26200123d900994a328b rd Xinqi Robotics Nexforce | Series A+ · hardware With participation from Fu Rande, Wuzhong Financial Holdings, Zhihui Investment, and the Ningbo Angel Guidance Fund. This marks the company's third round in six months since it began market-based fundraising in late 2025, building on Lixing Co.'s materials and transmission foundation to make integrated "materials-transmission-module" humanoid core actuation components. Source: First Robot source WeChat, CN https://mp.weixin.qq.com/s? biz=MzkwMTY1OTQ4OQ==&mid=2247484759&idx=5&sn=6c52094fb9ac02cf156de2dd8d7dc6c3&chksm=c1f7cbb84e67686a5b497ee997ad92e2d073256f3d3d6e087ad7f67f835a513351b0efb28201 rd Yunmu Zhizao Suzhou | Pre-A round | Tens of millions of yuan · humanoid Strategic investment from listed force-sensor maker Kelli Sensing . The team comes from an aerospace robotic-arm background, and this round coincides with the launch of a 60kg dual-arm heavy-load humanoid robot targeting manufacturing scenarios. The company states its revenue mix will evolve toward a "40/60 split" between industrial and cultural-tourism sectors. ⚠️ Company's own accountSource: Xinhua Daily Finance source WeChat, CN https://mp.weixin.qq.com/s? biz=Mzg5MDk2NDAyOQ==&mid=2247519054&idx=1&sn=6e751f8a3238781ba90e77dd0cf4cbe4&chksm=cef31e2e95ce23bb40708406b45796d710760cae9f7d0e42e4efe5388de1d28abbdc41895dd9 rd Oversonic Robotics Italy | STMicroelectronics takes a stake · humanoid STMicroelectronics confirmed a stake in this Italian humanoid robotics startup, whose RoBee humanoid targets industrial and medical scenarios. Funding will go toward technology platform development, manufacturing capacity, and production scale-up. European humanoid-robot companies are scarce, and a chip giant entering as an industrial investor is a concrete sign of Europe shifting from bystander to participant. Source: Asimov's Robots source WeChat, CN https://mp.weixin.qq.com/s? biz=MzY5NTA3MzczOA==&mid=2247485687&idx=1&sn=565f40bc091ca073d8add12a265ccf0c&chksm=f548652ad658037451ba0c88974298d8bf6cf7d9e7de2c8d79d0137476ed6b2af13d39074ac5 rd Yingzhi XBOT | Series B | 300-500 million yuan · adjacent Funded by Hong Kong's Jiankun Capital, with joint participation from government funds, US-dollar funds, and industry players, as one of two consecutive rounds. Makes general-purpose food-service embodied robots. On the same day, commercial coffee-robot company Kaye also secured funding — nearly 1 billion yuan combined flowing into food-service scenarios on the same day, suggesting capital is repricing narrow use cases where "humanoids can't do it, but someone's definitely willing to pay."Source: Smart Kitchen Talk source WeChat, CN https://mp.weixin.qq.com/s? biz=Mzg3OTU1NjQ5NQ==&mid=2247521761&idx=1&sn=bc0d0e5686be011686ce65e7a8e382b9&chksm=ce6d4ec8db5573327acf648928cf3a1c197a67bfe526d32e1d0d2b71f84e7c4f42548c9ea444 rd China Mobile procures 400 humanoid robots for 124 million yuan, split between Zhiyuan and Unitree · humanoid This is the largest publicly disclosed single humanoid robot order in China. According to the winning-bid announcement on China Mobile's procurement and tendering website, the full-size humanoid package was won by Zhiyuan Chinese humanoid robotics startup for 78 million yuan , covering 200 full-size bipedal robots plus 100 voice kits, 100 five-finger dexterous hands, 100 depth cameras, and 100 lidar units; the small-size humanoid and dexterous-hand package went to Unitree for 46.05 million yuan , also for 200 units. The key detail is buried in the procurement language — the announcement specifies "OEM service procurement," locking in three years of production capacity rather than a one-time purchase. The procuring entity, China Mobile Hangzhou Information Technology, is a wholly-owned subsidiary of China Mobile, and the stated purpose is to research key human-machine interaction technologies for general-purpose humanoid robots based on large language models: the carrier is buying the "body," while what it aims to keep for itself is the "brain" and the real-world data these 400 robots will generate. Zhiyuan's general business president Wang Chuang responded to delivery-pressure questions by saying "no pressure at all," with clients taking delivery in batches by use case. Viewed soberly, 400 units is small relative to both companies' planned capacity of tens of thousands of units, and the real test is whether these robots can run stably once deployed; but it moves humanoid procurement up a tier — from research institutions and automaker pilot lines to carrier-scale systematic procurement . Source: 51 Smart Choice Robots source WeChat, CN https://mp.weixin.qq.com/s? biz=MzcwOTA2NjU0Mw==&mid=2247484135&idx=1&sn=d9b71c6184b08840af0f8b7d8517940f&chksm=f42c8c4e6c316208403720476ac0ad248750398066e577c79973fdd4aeae4cadedb055dcf477 rd Galbot wins 236 million yuan bid for Yibin high-speed rail station, 500 robots set single-procurement record · embodied According to the bid-winner announcement for the "Embodied AI Robot Bodies and Related Equipment Procurement Project" by Yibin High-Speed Rail South District Hechuang Energy Management Co., Beijing Galbot Chinese embodied AI startup became the top candidate with a bid of 235.92 million yuan , for a procurement scope of 500 units of embodied AI robot bodies and supporting equipment, against a budget of 257 million yuan — setting a new record for the largest single embodied AI robot procurement since 2026. The wheeled robots will be used for high-speed rail station patrol, guidance, and cleaning, extending into logistics handling and retail terminals. Notably, the buyer's identity: a joint venture between the Yibin municipal urban investment platform and a new-energy platform under Wuliangye Group — i.e., a local state-owned industrial platform — showing the core buyers of embodied robots shifting from research institutions toward local government platforms and large state-owned enterprises. This is Galbot's second major order within a month having won a 39.17 million yuan project in Taizhou just in June . The company has already formed a closed loop in Yibin: last September its Southwest Embodied AI Training Center was established with 70 data-collection robots deployed, a local Yibin subsidiary was set up the same month, and it has now landed this large-scale procurement. The suspense lies in delivery — whether consistency can be maintained across scaled production within the 12-month delivery window remains an unverified engineering challenge. Source: Deep Current Sohu source https://m.sohu.com/a/1049686865 121608821?scm=10001.325 13-325 13.0.0-0-0-0-0.5 1334 Zhiyuan's 8 Genie G2 units handle an entire 3C quality-inspection line, running 17,000+ operations over 6 consecutive days · humanoid The Financial Times' front page covered Zhiyuan's Lingxi X2, framing it around China using humanoid robots to address manufacturing labor shortages. Deployment figures compiled in the report: at a 3C tablet contract manufacturer, 8 Genie G2 units handled a complete quality-inspection line, running 17,000+ operations over 6 consecutive days with a success rate of 99.99% ; at a new-energy vehicle assembly plant, units participated in mass-production assembly with repeat positioning accuracy of ±0.1mm. The significance of these figures lies in duration rather than peak performance — sustained multi-day production-line stability is exactly the line that separates "demo mode" from "deployment mode." ⚠️ Company's own accountSource: Zhongxing Zhiyun source WeChat, CN https://mp.weixin.qq.com/s? biz=MzYyNDI3ODI0Mg==&mid=2247484353&idx=1&sn=e99f4101aaf62eee426da4c225720f5a&chksm=f175de3354ef4b113e0a5e71243d8689186691615da7a8849654a1088ea762705506da414efa rd Xingdong Jiyuan Chinese humanoid robotics startup begins four-figure-unit deliveries in Q2, logistics center efficiency reaches 85% of human workers · humanoid The company states it began four-figure-unit batch robot deliveries in Q2 2026, with products already achieving routine operation at over a dozen logistics centers , reaching operational efficiency of 85% of human workers in some scenarios and running continuously 24 hours a day. This 85% figure is far more useful than vague phrases like "approaching human-level" — it means that in these scenarios, robots haven't yet crossed the replacement threshold, but have entered a range where the economics start to pencil out. ⚠️ Company's own accountSource: Robot Trend source WeChat, CN https://mp.weixin.qq.com/s? biz=MzkzNjkwNTg1NA==&mid=2247484634&idx=1&sn=bb59f37ecb08b1f12053b0783d9a4d2f&chksm=c3de9f9c54c386137c9b66fd041d14f1d5929aa603761f5db9a1606599d53b5cf565e8586e53 rd Baidu's Apollo Go enters South Korea, cumulative global rides surpass 20 million · autonomy Apollo Go is expanding into the South Korean market, while also disclosing that cumulative global rides have surpassed 20 million . On the same day, Hong Kong is accelerating autonomous driving testing, and Baidu is expanding its pilot program there. China's robotaxi export push is timed right against tightening US regulation see the top industry-dynamics item — right now, the overseas compliance window is worth more than the technology window.Source: Autohome source https://chejiahao.autohome.com.cn/info/24886071 Komatsu's autonomous mining trucks surpass 1,000 units deployed · industrial Komatsu's fleet of driverless mining trucks has crossed 1,000 units . Another aligned development comes from Western Australia — a fleet of retrofitted HD1500 autonomous trucks has begun unmanned day-shift operations. Mines were among the earliest closed environments where autonomous driving economics worked out, and "retrofitting existing trucks" is a far cheaper path than replacing entire vehicle fleets.Source: TechnoSports source https://technosports.co.in/komatsu-hits-1000-autonomous-mining-titans/web-story/ JAKA Robotics and StarCharge form strategic partnership, bringing embodied robots to EV charging scenarios · embodied JAKA's embodied AI robots will enter StarCharge's charging-station environments. Automated charging is a rare combination for manipulation robots — "clearly bounded task + high repetition + a job people don't want to do." The precision demands of plug alignment are real, but the environment is far more structured than a home setting.Source: Guandian.cn source https://www.guandian.cn/article/20260713/573055.html US NHTSA issues ultimatum to all robotaxi operators: submit remediation plans for emergency-response obstruction by month's end · autonomy National Highway Traffic Safety Administration head Jonathan Morrison issued a formal directive to every autonomous vehicle developer on the Department of Transportation's Standing General Order list, demanding a solution to the problem of "robotaxis obstructing emergency responders," with a deadline of the end of this month . The letter named no specific company, but the trigger is clear: during the July 4th Golden Gate Bridge fireworks night, San Francisco's north corridor was jammed with crowds, and multiple Waymo vehicles got stuck in the congestion, several of which had to be towed after their batteries ran out while idling. San Francisco supervisor Bilal Mahmood has announced he will send letters demanding that the city's emergency management department, fire department, and transportation authority explain the impact of autonomous vehicles on traffic and emergency response that night. The real weight of this directive lies in its reclassification — it moves "responding to emergency situations" from an edge case to a baseline capability , thereby redefining what counts as a compliant autonomous driving system, with compliance costs falling directly on operators who have already deployed at scale. Intriguingly, another signal emerged the same week: the 2026 regulatory agenda proposes amending FMVSS to clear obstacles for vehicles without steering wheels or pedals the design direction Tesla and Zoox are pursuing . Tightening behavioral requirements on already-deployed fleets while loosening design constraints for future vehicle models — these two opposing federal signals stacked together favor well-capitalized players who can absorb short-term compliance costs and also have the lobbying capacity for long-term hardware deregulation. The same day saw Uber's robotaxi lobbying clash directly with Waymo, and the Washington D.C. Council is also set to debate autonomous vehicle legislation.Source: Silicon Canals source https://siliconcanals.com/sc-n-the-federal-regulator-just-told-every-robotaxi-operator-to-fix-emergency-response-failures-by-months-end-and-the-quieter-rule-change-happening-alongside-it-reveals-who-actually-wins/ ByteDance explores autonomous driving, led by its Seed world model team · world-model 36Kr exclusive: ByteDance is exploring entry into autonomous driving, with the project led by Zhou Chang 's world model team under the Seed organization, aimed at driverless logistics and reporting into Volcano Engine's automotive industry line. Zhou Chang joined ByteDance in 2024 to lead multimodal large models, and his scope has steadily expanded — the visual generation business was folded in last year, and this year the Seed Robotics team now reports directly to him as well, covering multimodal large models, world models, visual generation, and embodied AI. ByteDance has previously had business contact with leading autonomous driving teams and has extended offers to multiple autonomous-driving talent. The company's response: it has substantial early-stage research in frontier directions like physical AI, but has no plans for an autonomous driving business . Technically, this path is self-consistent — world models need to understand the physical world and predict future states, which is exactly the capability autonomous driving requires; this year, Geely, Momenta, Xiaomi, Qingzhou Zhixing and others have progressively rallied under world models, including some former VLA-camp players. ⚠️ Media report, company deniesSource: 36Kr source https://eu.36kr.com/en/p/3893815451417347 SoftBank × Yaskawa Electric × NVIDIA: running the full Physical AI pipeline on GPU cloud, demonstrating deformable object manipulation · world-model SoftBank, together with Yaskawa Electric and NVIDIA, used SoftBank's under-construction "AI data center GPU cloud" as a Physical AI development platform to demonstrate a deformable object manipulation system developed by Yaskawa. Traditional robot control typically assumes the target object's shape and position are known, making flexible objects like ropes, cloth, and bags — whose form changes every time — a persistent challenge; Yaskawa uses VLA to generate actions directly from vision and task instructions to identify, grasp, and manipulate them. The platform integrates the NVIDIA Omniverse simulation library, the Cosmos open-world model, and the Physical AI Factory Blueprint, connecting action data collection → model training → simulation evaluation → real-robot deployment into a single pipeline. SoftBank's goal isn't to sell robots, but to lower the barrier to Physical AI development so that "engineers who aren't robotics experts can use it" — a typical positioning move for cloud vendors in the embodied AI space.Source: SoftBank official source https://www.softbank.jp/en/corp/news/press/sbkk/2026/20260713 01/ Tesla tears down Model S/X production line to build Optimus, AI lead says "there are big shoes to fill" · humanoid Tesla has torn down the Model S/X production line at its Fremont factory, converting the freed-up space into an Optimus production line. AI lead Ashok Elluswamy commented that Optimus has "big shoes to fill" — swapping out a production line that once built a flagship car model for humanoid robot capacity says more about Tesla's resource prioritization than any demo could.⚠️ Statement per companySource: InsideEVs source https://insideevs.com/news/801412/tesla-model-s-model-x-factory-teardown/ Samsung obtains design blueprints for Tesla's AI5 autonomous driving chip, mass production set for next year · hardware Samsung has obtained design blueprints for Tesla's AI5 autonomous driving chip, with mass production planned for next year. AI5 is the next-generation compute base shared by Tesla's FSD and Optimus; foundry work landing at Samsung means this core Physical AI supply chain is further diversifying away from sole reliance on TSMC.Source: Autohome source https://chejiahao.autohome.com.cn/info/25963171?reply=reply LG Electronics establishes a robotics business center, elevating robotics into an independent growth engine · industrial LG Electronics has set up a new Robotics Business Center covering home, office, and factory scenarios. This reads more clearly alongside two other South Korea items from the same day: the South Korean government has pledged full fiscal and infrastructure support for semiconductors and AI robotics and will fast-track related legislation; and Holiday Robotics' Series A included direct investment from two policy banks. Conglomerates, startups, and state capital are all mobilizing simultaneously.Source: Passionate In Marketing source https://www.passionateinmarketing.com/lg-electronics-establishes-robotics-business-center-to-accelerate-future-growth/ Hyundai Motor union goes on strike, one demand being job protection against "robots taking jobs" · industrial Hyundai Motor workers have gone on strike over bonuses and robot-related job security issues. This is a rarely-seen flip side of the industry narrative — while humanoid and industrial robots are counted as capacity and valuation, production-line workers are writing them into collective bargaining terms. Germany's industrial sector reportedly lost 124,000 jobs in 2025, with humanoid robots cited as one contributing factor. This kind of friction will appear increasingly often along the deployment path.Source: NDTV Profit source https://www.ndtvprofit.com/business/hyundai-motor-workers-strike-over-bonuses-robot-job-security-11764190 Guangdong releases industry-group standard for full-chain reliability control of industrial robots · industrial Guangdong Province has released the "Guide to Reliability Control and Quality Improvement for the Industrial Robot Industry Chain," described as filling a gap in the industry — extending reliability requirements from the finished machine to the entire supply chain . China's robotics industry policy lever is shifting from "subsidizing capacity, hosting competitions and expos" toward "setting quality benchmarks," which represents a real compliance cost for upstream component makers and a filter squeezing out low-price, low-quality capacity.Source: 21jingji source https://m.21jingji.com/article/20260713/herald/4493b7ca2fb5d64ee21ce3ab5784c069.html Shangwei New Materials unveils Qiyuan T1: automatic switching between wheeled-humanoid and quadruped forms · humanoid The Qiyuan T1 will debut at WAIC 2026, based on a single body that achieves automatic shape-shifting between wheeled-humanoid and quadruped forms , targeting home consumer scenarios. Breaking away from a fixed form factor is an interesting design direction — but a shape-shifting structure inevitably comes at the cost of more joints, added weight, and reliability; a demo switch on the show floor is a different matter from running for three months at home.⚠️ Demo per companySource: Embodied Home Robots source WeChat, CN https://mp.weixin.qq.com/s? biz=MzUzNjEwMjg4Nw==&mid=2247522675&idx=1&sn=84107cb0c40b13025f17ac97d36b6adb&chksm=fbba6a0cafa302ef4da096fadfd2a82ddb1d2ad79ffd58f1f18a57e28a9a605507e169f94345 rd · Midea Welling · Reducer capacity : signed an agreement with the Shunde District government in Foshan to build a new manufacturing base, planning 55,000 sqm of temperature-controlled, vibration-isolated precision production floor, with Phase 1 online in 2027 and Phase 2 in 2030, adding 300,000 units of reducer capacity . For reference, Midea shipped over 30,000 precision reducers and over 70,000 servo motors in 2025 — this is a roughly tenfold capacity pre-buildout source https://www.163.com/dy/article/L1ODNV3S05568W0A.html · LinkerBot · Dexterous hand capacity and pricing : peak monthly output over 4,000 units , 2025 revenue of 260 million yuan, and over 400 million yuan in orders on hand as of early 2026; according to Goldman Sachs research, its products are 10,000-20,000 yuan cheaper than competitors, with ASP 50% lower than overseas products. The company is targeting a Hong Kong IPO at a valuation of roughly 40 billion yuan, which has already drawn questions about "a valuation larger than the market itself" ⚠️ Company's own account source WeChat, CN https://mp.weixin.qq.com/s? biz=MzkyNDIxMDQ1OA==&mid=2247503988&idx=1&sn=57e21d17b95956c3b944ac342d96b722&chksm=c0518702a2a2867a79ab83ade14bfeadb3800f0af15d203b5de5b7696411fe1f837f77f08468 rd · DEEPX × Avnet : South Korean robotics AI chip company DEEPX has signed a distribution agreement with global distributor Avnet, using its channels to expand into the Asia-Pacific market — edge inference chips are starting to move through standard-component distribution channels, a signal that this category is heading toward mass production source https://en.prnasia.com/releases/apac/deepx-expands-apac-footprint-through-distribution-agreement-with-global-technology-distributor-avnet-540505.shtml